Abstract

   Synthetic lethality (SL) is widely used to discover the anti-cancer
   drug targets. However, the identification of SL interactions through
   wet experiments is costly and inefficient. Hence, the development of
   efficient and high-accuracy computational methods for SL interactions
   prediction is of great significance. In this study, we propose MPASL, a
   multi-perspective learning knowledge graph attention network to enhance
   synthetic lethality prediction. MPASL utilizes knowledge graph
   hierarchy propagation to explore multi-source neighbor nodes related to
   genes. The knowledge graph ripple propagation expands gene
   representations through existing gene SL preference sets. MPASL can
   learn the gene representations from both gene-entity perspective and
   entity-entity perspective. Specifically, based on the aggregation
   method, we learn to obtain gene-oriented entity embeddings. Then, the
   gene representations are refined by comparing the various layer-wise
   neighborhood features of entities using the discrepancy contrastive
   technique. Finally, the learned gene representation is applied in SL
   prediction. Experimental results demonstrated that MPASL outperforms
   several state-of-the-art methods. Additionally, case studies have
   validated the effectiveness of MPASL in identifying SL interactions
   between genes.

   Keywords: synthetic lethality prediction, knowledge graph,
   multi-perspective learning, deep learning, attention mechanism

1 Introduction

   Cancer is a genetic disease caused by the accumulation of multiple
   mutations resulting from the interaction of internal and external
   factors ([38]Barabási et al., 2011). Traditional cancer treatments such
   as chemotherapy often have serious side effects and harm healthy cells
   ([39]Hanahan and Weinberg, 2011). Synthetic lethality (SL) is a genetic
   interaction that kills cancer cells selectively without damaging
   healthy cells ([40]Boone et al., 2007; [41]Hartwell et al., 1997;
   Iglehart and Silver, 2009). SL offers a tremendous depth of research
   opportunities for anti-cancer drug development and targeted cancer
   therapy, with researchers making great efforts to identify SL pairs.
   Discovering SL gene pairs relies heavily on high-throughput wet-lab
   screening techniques including RNAi screening ([42]Bartz et al., 2006;
   [43]Luo et al., 2009; [44]Gregory et al., 2010; [45]Blank et al., 2013;
   [46]Chang et al., 2016) and CRISPR screening ([47]Han et al., 2017;
   [48]Shen et al., 2017). However, lab experiment-based screening methods
   are time-consuming and expensive and increase the risk of off-target
   effects ([49]Liu et al., 2019). Thus, there is an urgent need for
   efficient and economical methods to overcome the deficiencies of
   high-throughput screening techniques ([50]Huang et al., 2019).

   To overcome these limitations, a several computational methods have
   been developed for SL prediction. These methods fall into two
   categories: (i) knowledge-based methods and (ii) supervised
   machine-learning methods ([51]Zhu et al., 2023). Knowledge-based
   methods rely on prior knowledge or assumptions (i.e., gene mutations
   ([52]Lu et al., 2020) or CNVs ([53]Lu et al., 2018)) to detect SL
   pairs. For example, Zhang et al. ([54]Zhang et al., 2015) proposed a
   combination of data-driven models with signaling pathway knowledge to
   discover SL interaction pairs by simulating the effects of gene
   knockout on cell death. Srihari et al. ([55]Srihari et al., 2015) used
   copy-number and gene expression data to identify SL interactions.
   However, knowledge-based methods do not comprehensively utilize
   underlying patterns of known SL interactions. Machine learning methods
   such as decision trees ([56]Wong et al., 2004), support vector machines
   ([57]Paladugu et al., 2008; [58]Qi et al., 2008), random forests
   ([59]Das et al., 2019), and ensemble classifiers ([60]Pandey et al.,
   2010; [61]Wu et al., 2014) expedite the identification of SL pairs are
   challenging to apply to large-scale data due to the complex matrix
   operations.

   Tremendous developments in deep learning-based methods have shown them
   to be effective in many biomedical tasks, including drug-target
   prediction ([62]Mohamed et al., 2020), drug-disease prediction ([63]Yu
   et al., 2021) and drug synergy prediction ([64]Zhang et al., 2023)
   along with successful applications in SL prediction ([65]Huang et al.,
   2019; [66]Liu et al., 2019; [67]Cai et al., 2020; [68]Liany et al.,
   2020; [69]Hao et al., 2021; [70]Long et al., 2021). For example, Long
   et al. ([71]Long et al., 2021) proposed a graph contextualized
   attention network to predict SL interactions. This model deploys a
   dual-attention mechanism to capture the importance of neighbors and
   feature graphs for node representation learning. Cai et al. ([72]Cai et
   al., 2020) modeled SL interactions as a graph and adopted a dual-drop
   GNN to address the sparsity of SL networks. However, most of these
   methods are limited in the expressive capacity of homogeneous graphs.

   Knowledge graphs (KGs) are multi-relational heterogeneous graphs where
   the nodes and edges correspond to different types of entities and
   relations, respectively ([73]Wang et al., 2017; [74]2019b). They
   overcome the limitations of homogeneous graphs by using rich semantic
   information between graph entities to discover potential relations.
   These have begun to equip bioinformaticians with powerful weapons for
   combining heterogeneous data plainly for SL pairs prediction. Wang et
   al. ([75]Wang et al., 2021) presented a KGNN-based model, KG4SL, to
   predict SL interactions. It uses independent knowledge embeddings to
   capture the underlying biological mechanisms of interconnected SL
   pairs. Zhu et al. ([76]Zhu et al., 2023) utilized relations in
   knowledge graphs to represent SL-related factors and learned latent
   representations of genes through message aggregation. It is evident
   that employing KG entities such as gene, pathway and their neighbors
   yields a more accurate embedding representation, but previously
   KG-based methods ignore the preferences of existing SL interactions and